Assembly Integrity Detection Based on 3D CAD Model

Abstract

Incomplete assembly could affect product quality and performance, and even cause inestimable losses. In the industrial environment, the lack of image datasets and the inability to obtain the physical assembly in advance is a challenge for machine vision to detect assembly integrity. We propose a 3D CAD model-based method to detect assembly integrity without physical assembly. 3D CAD models are utilized for 2D image rendering and dataset construction, and automatic labeling can be achieved through edge extraction and minimum enclosing rectangle fitting rather than by hand. An end-to-end detection neural network based on Faster-RCNN is trained in the datasets. The VGG network as detection neural network to extract features, and the generated feature map determines the candidate detection region through RPN. Fast-RCNN then detects the object in the image. Finally, by detecting images from multiple views with dimension reduction and comparing them with prior knowledge, we can judge whether there are missing parts in the assembly. Experimental results show that the network model can accurately detect the integrity of the assembly, and the average accuracy (mAP) reaches 88.9%. The method can also provide theoretical guidance for future physical assembly and shorten the experimental process.

Keywords

• Assembly integrity detection
• convolutional neural network
• CAD model
• Faster-RCNN